A road surface collapse identification and detection method based on a YOLOv5 network model

Abstract

The application relates to the field of computer image detection and recognition, and provides a road surface collapse identification and detection method based on a YOLOv5 network model. The purpose is to provide a solution to the problem that if the model processing is not timely when the sample size is too large, the real-time identification function cannot be achieved. The main scheme comprises the following steps: obtaining an underground plane scanning graph containing time sequence characteristics and a profile real-time waveform graph containing time sequence characteristics; obtaining a region screening graph A by circling the plane scanning graph; obtaining a region screening graph B by using a buffer scanning area-main scanning area-buffer scanning area structure on the profile real-time waveform graph; combining the region screening graph A and the region screening graph B into a data pair according to the time sequence characteristics; labeling each data pair with a geological disaster category, and then dividing the data pairs into a test set and a training set; completing YOLOv5 model training through the training set and the test set; obtaining a geological radar graph in real time, processing the data pair, and then sending the data pair into the trained YOLOv5 model for geological disaster category identification.

Description

TECHNICAL FIELD

[0001] The application relates to the field of computer image detection and recognition, and provides a road surface collapse identification and detection method based on a YOLOv5 network model. BACKGROUND

[0002] At present, there are two processes for identifying the types of road surface collapse disasters. The first process is to obtain the area where the risk is likely to occur from the geological radar graph. The geological radar graph under the road surface is obtained by detecting the road surface by using a geological radar detector, and then a geological disaster identification personnel with very rich experience circles out the area where the risk is likely to occur by using manpower. The second process is to analyze the cross-section graph corresponding to the area where the risk is likely to occur. The cross-section graph corresponds to a wave inside the land. The wave is analyzed by the disaster identification personnel to determine whether the disaster occurs at the current position, and further determine which disaster occurs, and finally a conclusion is drawn.

[0003] The above-mentioned current road surface collapse disaster type identification process has three obvious defects. First, the professional ability of the disaster identification personnel needs to be very strong. When the address radar graph cross-section area is circled, the identification personnel is required to have rich experience and to be able to subjectively combine objective factors to make correct judgments on a certain area. Otherwise, if too little is circled, a very serious disaster area may be missed; if too much is circled, too much time cost will be wasted to analyze unnecessary areas. Second, the time cost is too high. In the process of identifying geological disasters, the road surface is first detected by using a geological radar detector to obtain a geological radar graph, and then the area is circled by the identification personnel one by one for analysis. For the detection of the entire road surface, it is often several tens or even hundreds of kilometers. If the personnel only judge one by one, it can be imagined that the time cost of the entire process is very high. Third, although the professional personnel can circle the corresponding area where the risk is likely to occur, the accuracy rate is different for different people, and there is no way to guarantee a high accuracy rate.

[0004] Real-time identification of underground collapse can report the problems of the detected road section at the fastest speed, so that the hidden danger can be reported at the first time. This is the problem to be solved by the project. However, in the current scene, the position of the area where the underground collapse occurs is not fixed, and the length may be much larger than the length of the vehicle body or the range supported by the detector, which leads to the fact that the existing image area splicing method needs to process a large amount of image data, so that the image block splicing method may not be applicable. SUMMARY

[0005] The purpose of the present application is to provide a solution to the problem that if the model processing is not timely when the sample size is too large, the real-time identification function cannot be achieved.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical means:

[0007] A road surface collapse identification and detection method based on a YOLOv5 network model, comprising the following steps:

[0008] Step 1: Obtain a geological radar graph as a sample, which is a ground plane scanning graph containing time sequence characteristics and a profile real-time waveform graph containing time sequence characteristics, and two images are obtained in horizontal and vertical directions using double signal sources;

[0009] Step 2: Perform a region operation on the gray value of the plane scanning graph to obtain a region that may have risks, denoted as region A, and divide region A on the plane scanning graph to obtain a region screening graph A containing only region A by removing the region without risks;

[0010] Step 3: For the profile real-time waveform graph, use a buffer scanning area-main scanning area-buffer scanning area structure to obtain a region B composed of boundary-collapse boundary information, and divide region B on the plane scanning graph to obtain a region screening graph B containing only region B by removing the region without risks;

[0011] Step 4: Form a data pair according to the time sequence characteristics of the region screening graph A and the region screening graph B;

[0012] Step 5: After labeling each data pair with a geological disaster category, divide the test set and the training set;

[0013] Step 6: Complete the YOLOv5 model training through the training set and the test set, and obtain the trained YOLOv5 model;

[0014] Step 7: Real-time obtain the geological radar graph, process it through steps 2-4, and input it into the trained YOLOv5 model for geological disaster category identification.

[0015] In the above technical solution, step 2 specifically includes:

[0016] Step 2.1: The plane scanning graph is referred to as a sample picture, and an average analysis of the sample picture obtains a gray value N. The first traversal of the sample picture is performed from the first pixel of the sample picture, and pixels greater than N are set to 1 and pixels less than N are set to 0 from left to right. After the first traversal, the sample picture is divided into regions containing only 0 or 1, and a fragmented graph is obtained.

[0017] Step 2.2: A second traversal is performed, and each pixel point in the fragmentation graph obtained in step 1.1 is traversed in turn. If the gray value is 0, it is skipped; if the gray value is 1, the eight nearest neighbor pixel points around the pixel point are traversed, and the pixel points with a value of 1 are added to a set. After the traversal is completed, the entire region is added to a set, and the relative position of the region in the picture is obtained. Then, the next part of the pixel points is continuously traversed until the entire image is traversed. At this point, n sets are obtained, and the n sets represent n regions that meet the requirements. Finally, the regions are separated on the picture to obtain a screening graph A that only contains the regions that may have risks and excludes the blank regions.

[0018] In the technical solution, step 3 specifically includes the following steps:

[0019] For the profile real-time waveform graph, a main scanning section with a width of Mpx and a length equal to the height of the profile real-time waveform graph is used to scan the profile real-time waveform graph in the time sequence increasing direction. On both sides of the main scanning section, a buffer scanning area with a width of Npx is bound, and N is a custom parameter, which is used to detect whether there is a boundary.

[0020] For the main scanning section, if the absolute value of the amplitude determination value in the profile real-time waveform graph is greater than the threshold value, the result is a collapse area.

[0021] For the buffer scanning area, after calculating the Fourier transform result of the buffer area of the profile real-time waveform graph in the width direction, if the change amplitude is greater than the threshold value, it is marked as a boundary.

[0022] Through the buffer scanning area, the main scanning area, and the buffer scanning area, a region B composed of a boundary-collapse area-boundary is obtained. The region B is separated on the plane scanning graph to obtain a region screening graph B that only contains the region B and excludes the regions without risks.

[0023] Because the application adopts the above technical solution, the following beneficial effects are achieved:

[0024] The application has the advantages of being able to identify the feature parts of the plane graph and the profile graph at the same time through steps 2-4 with small hardware overhead, find these possible problem areas in real time, and then classify and identify them through a model. This approach greatly reduces the input amount of model samples, improves the identification efficiency of the model, and uses the combination of the underground plane scanning graph containing time sequence characteristics and the profile real-time waveform graph containing time sequence characteristics to input the model for identification, which can verify each other and improve the identification accuracy.

[0025] In the case of single identification, the plan view cannot accurately identify the longitudinal data and can only identify the layered image of the road plane; the profile view cannot identify the area of each layer of the plane and can only identify the waveform perpendicular to the plane. Therefore, only by combining the two types of graphs can the area where problems may exist be accurately found and its three-dimensional coordinates be located. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 For a cavity type collapse imaging graph;

[0027] Figure 2 For a loose type collapse imaging graph;

[0028] Figure 1 And Figure 2 The figure shows the image conditions of two types of road surface collapse, the upper half of the image is a plan view, and the lower half is a vertical profile view. As can be seen, the collapsed area presents a continuous dark area in the original detection graph (here, orange is used for marking, and the original image is a continuous dark area). Using the scheme proposed in this project, such parts can be effectively identified, and the area of the image is automatically marked. DETAILED DESCRIPTION

[0029] The embodiments of the present application will be described in detail below. Although the present application will be described and illustrated by combining some specific embodiments, it should be noted that the present application is not limited to only these embodiments. On the contrary, modifications or equivalent replacements of the present application should be covered in the scope of the claims of the present application.

[0030] In addition, in order to better illustrate the present application, numerous specific details are given in the specific embodiments below. Those skilled in the art will understand that the present application can also be implemented without these specific details.

[0031] Example 1

[0032] The present application provides a road surface collapse identification and detection method based on a YOLOv5 network model, comprising the following steps:

[0033] Step 1, obtain a geological radar graph as a sample, the geological radar graph is an underground plan scan graph containing time sequence characteristics and a profile real-time waveform graph containing time sequence characteristics, two images are obtained in horizontal and vertical directions using double signal sources;

[0034] Step 2, perform a region operation on the gray value of the plan scan graph to obtain a region where a risk may exist, denoted as region A, and divide region A on the plan scan graph to obtain a region screening graph A containing only region A by removing the region where there is no risk;

[0035] Step 3, for the cross-section real-time waveform chart, using the buffer scanning area-main scanning area-buffer scanning area structure to obtain the area B composed of boundary-collapse boundary information, the area B is divided on the plane scanning chart to obtain the area screening chart B containing only the area B without the risk area;

[0036] Step 4, the area screening chart A and the area screening chart B are combined according to the time sequence characteristics to form a data pair;

[0037] Step 5, after labeling each data pair with a geological disaster category, dividing the test set and the training set;

[0038] Step 6, training the YOLOv5 model through the training set and the test set to obtain the trained YOLOv5 model;

[0039] Step 7, real-time acquisition of the geological radar chart, after processing by steps 2-4, inputting into the trained YOLOv5 model for geological disaster category identification.

[0040] In the above technical solution, in the original image, the potential collapse of the road surface is marked by the dark area, and the non-collapse area is marked by the light area. The brightness of the dark area in the image is different, the size is different, and the shape is different, so it is necessary to use the program to check and verify during the initial labeling data. For this step 2 specifically includes:

[0041] Step 2.1: the plane scanning chart is referred to as a sample picture, and a gray value N is obtained by averaging the sample picture. The first time the sample picture is traversed, it is judged from the first pixel of the sample picture to the right. If the gray value is greater than N, it is set to 1, and if the gray value is less than N, it is set to 0. After the first traversal, the sample picture will be divided into a region containing only 0 or 1, and a fragmentation chart is obtained;

[0042] Step 2.2: the second traversal is performed, and each pixel point in the fragmentation chart obtained in step 1.1 is traversed in turn. If the gray value is 0, it is skipped, and if the gray value is 1, the 8 nearest neighbor pixel points around the pixel point are traversed. The pixel point with a value of 1 is added to a set. After the final traversal, the region is added to a set, and the relative position of the region in the picture is obtained. Then continue to traverse the next part of the pixel point until the entire image is traversed. At this time, n sets are obtained, and the n sets represent n regions that meet the requirements. Finally, the region is divided on the picture to obtain the area screening chart A containing only the possible risk area.

[0043] The above technical solution, step 3 specifically includes the following steps:

[0044] For the profile real-time waveform chart, a main scanning section with a width of 60px and a length equal to the height of the profile real-time waveform chart is used to scan the profile real-time waveform chart in the time-increasing direction, and a buffer scanning area with a width of 20px is bound on both sides of the main scanning section to detect whether there is a boundary; in general, for the main scanning area, due to the complex collapse area and uneven soil, the result with a larger absolute value of amplitude determination value is defined as the collapse area, and for the buffer scanning area, the purpose is to detect the boundary, so after calculating the Fourier transform of the function image in the buffer area in the width direction, if the change amplitude is found to be too large, it is marked as a boundary and matched with the boundary verified in the horizontal direction. When the matching is successful, it can be determined as a suspected collapse boundary, and a region is circled in the horizontal direction. The other results in the vertical direction generate a value between 0 and 1 as a suspected collapse score according to whether it is a collapse, and the closer to 1, the more likely it is a collapse area.

[0045] For the main scanning area, the result with an absolute value of amplitude determination value greater than the threshold value in the profile real-time waveform chart is a collapse area;

[0046] For the buffer scanning area, after calculating the Fourier transform of the buffer area of the profile real-time waveform chart in the width direction, if the change amplitude is greater than the threshold value, it is marked as a boundary;

[0047] Through the buffer scanning area, the main scanning area, and the buffer scanning area, a region B composed of a boundary-collapse area-boundary is obtained, and the region B is divided on the plane scanning chart to obtain a region screening chart B containing only the region B without the risk area.

[0048] After the above processing is completed, a region screening chart B containing a suspected collapse area with complete boundaries and a region screening chart A are obtained, and then input into the YoloV5 model for prediction. Due to the single scene and high quality of the preprocessed data, the accuracy of the test set can reach 97.52%. In the actual running process, the images collected by the car are as shown in the figure, the boundary information is automatically circled, and the horizontal image is predicted using the model to obtain the final prediction result.

Claims

1. A road surface collapse identification and detection method based on a YOLOv5 network model, characterized in that, The method comprises the following steps: Step 1, obtaining a geological radar chart as a sample, the geological radar chart being a ground plane scanning chart containing time sequence characteristics and a profile real-time waveform chart containing time sequence characteristics; Step 2, performing a region operation on the gray value of the plane scanning chart to obtain a region possibly containing risks, denoted as region A, and segmenting region A on the plane scanning chart to obtain a region screening chart A containing only region A and removing regions without risks; Step 3, using a buffer scanning area-main scanning area-buffer scanning area structure to obtain region B composed of boundary-collapse area-boundary information, and segmenting region B on the plane scanning chart to obtain a region screening chart B containing only region B and removing regions without risks; Step 4, grouping region screening chart A and region screening chart B into data pairs according to time sequence characteristics; Step 5, labeling each data pair with a geological disaster category, and dividing the data pairs into a test set and a training set; Step 6, training a YOLOv5 model through the training set and the test set, and obtaining the trained YOLOv5 model; Step 7, obtaining a geological radar chart in real time, processing the geological radar chart through steps 2-4, and inputting the processed geological radar chart into the trained YOLOv5 model for geological disaster category identification; Step 2 specifically comprises: Step 2.1: the plane scanning chart is referred to as a sample picture, a gray value N is obtained by averaging the sample picture, the sample picture is traversed for the first time, and a first pixel of the sample picture is judged first, pixels greater than the gray value N are set to 1 from left to right, and pixels less than the gray value N are set to 0, after the first traversal, the sample picture is divided into regions containing only 0 or 1, and a fragmentation chart is obtained; Step 2.2: the fragmentation chart obtained in step 1.1 is traversed for the second time, if the gray value of each pixel point in the fragmentation chart is 0, the pixel point is skipped, if the gray value of each pixel point is 1, the eight nearest neighbor pixel points around the pixel point are traversed, the pixel points with a value of 1 are added to a set, after the traversal, the region is added to a set, the relative position of the region in the picture is obtained, then the next part of the pixel points is continuously traversed, until the entire image is traversed, at this time, n sets are obtained, and the n sets represent n regions meeting the requirements, finally, the regions are segmented on the picture to obtain a region screening chart A containing only the regions possibly containing risks and removing blank regions.

2. The road collapse identification and detection method based on the YOLOv5 network model according to claim 1, characterized in that: Step 3 specifically comprises the following steps: The profile real-time waveform chart is scanned in the time sequence increasing direction using a main scanning area section with a width of Mpx and a length equal to the height of the profile real-time waveform chart, and a buffer scanning area with a width of Npx is bound on both sides of the main scanning area section, N being a self-defined parameter, for detecting whether a boundary exists, For the main scanning area, if the absolute value of the amplitude determination value in the profile real-time waveform chart is greater than a threshold value, the result is a collapse area; For the buffer scanning area, the result of the Fourier transform of the buffer area of the profile real-time waveform chart in the width direction is calculated, and if the change amplitude is greater than a threshold value, the result is marked as a boundary. The region B composed of the buffer scanning area, the main scanning area, the buffer scanning area and the border-collapsed area-border is obtained, and the region B is divided on the planar scanning graph to obtain a region screening graph B containing only the region B without the risk-free region.

Images